feat: wire invoke_capability() dispatch, CLI run command, and integration tests

[marcuspat]

Summary

Production wiring that connects the reasoning loop to real capability implementations and turns the sentinel run CLI command into a working investigate → plan → approve → act flow.

Reasoning loop (sentinel-agent-llm/src/reasoning_loop.rs)

• Added a capability_impls: HashMap<String, Box<dyn Capability>> field plus a with_capabilities(...) builder.
• invoke_capability() now dispatches to the registered concrete capability via cap.invoke(args, ctx), falling back to the previous stub result when no implementation is registered (preserving existing stub-mode behavior/tests).
• Rollback path now calls cap.invoke_inverse(...) for completed steps before marking them RolledBack, logging success / failure / no-inverse.

CLI (sentinel-tui/src/main.rs)

• New run_agent() wires the LLM backend (Anthropic/OpenAI), executor, all capabilities, registry, default policy, and audit log into a ReasoningLoop, then runs investigate → plan → approve → act. Approval comes from stdin or --auto-approve; --dry-run prints the plan without executing.
• Commands::Run gained --auto-approve; the stub body now calls run_agent(...).
• list_capabilities() now enumerates the real capabilities from all_capabilities(...) with kind, risk tier, and rollback support.

Tests (sentinel-agent-llm/tests/integration_test.rs)

• New end-to-end integration suite wiring real capabilities through a recording mock executor and a sequential mock LLM backend, covering: real capability dispatch during investigation, full investigate→plan→execute, policy-deny observation, audit-chain validity, the 14-capability registry, and stub-mode fallback.

Notes on deviations from the patch spec

A few surgical adjustments were required so the workspace compiles and cargo test --workspace passes:

• all_capabilities(...) requires an Arc<dyn CommandExecutorTrait>. The struct CommandExecutor does not implement that trait (only RealCommandExecutor does), so the executor used in both run_agent and list_capabilities is RealCommandExecutor. Added sentinel-exec to the TUI dependencies.
• The test's CommandOutput construction was corrected to the actual struct (exit_code: Option<i32>, plus the truncated field).
• CapabilityResult::Success exposes its payload as output (not data), and AuditLog exposes event_count() and persists on append (no separate flush()), so the test was adjusted accordingly.

Testing

• cargo build --workspace — clean
• cargo test --workspace — all green (integration suite: 6/6)
• cargo clippy --workspace --all-targets — no warnings

Do not merge.

https://claude.ai/code/session_01PrvrHsArvuyDmDvn5Cuvri

---

Generated by Claude Code